-
Notifications
You must be signed in to change notification settings - Fork 0
/
rnn.py
285 lines (194 loc) · 8.89 KB
/
rnn.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
#!/usr/bin/env python
# -*- coding: utf-8 -*-
# Author: Mathias Müller / mmueller@cl.uzh.ch
import numpy as np
from typing import Sequence
from collections import namedtuple
from activations import sigmoid, tanh
class RNN:
def __init__(self) -> None:
self.params = {}
def step(self, input_):
raise NotImplementedError
def forward_sequence(self, sequence):
raise NotImplementedError
class ElmanScalarRNN(RNN):
"""
An Elman or "vanilla" RNN that, at each time step,
takes scalar input.
>>> e = ElmanScalarRNN(10)
>>> print(e.step(5))
[0.48501544 0.49107564 0.504578 0.50550117 0.51805654 0.48205011
0.51270459 0.47850066 0.49718829 0.50798906]
"""
def __init__(self, hidden_size: int) -> None:
super().__init__()
# in this scalar input version, the weight matrices are
# vectors, and the biases are scalars
self.params["Wh"] = np.random.randn(hidden_size,) * 0.01
self.params["Uh"] = np.random.randn(hidden_size,) * 0.01
self.params["bh"] = np.zeros(1,)
self.reset_state()
def reset_state(self) -> None:
self.hidden_state = 0.0
def step(self, input_) -> np.ndarray:
input_linear = np.dot(self.params["Wh"], input_)
hidden_linear = np.dot(self.params["Uh"], self.hidden_state) + self.params["bh"]
# compute new hidden state
self.hidden_state = sigmoid(input_linear + hidden_linear)
return self.hidden_state
def forward_sequence(self, sequence: Sequence[np.ndarray]) -> Sequence[np.ndarray]:
self.reset_state()
hidden_states = []
for item in sequence:
hidden_states.append(self.step(item))
return hidden_states
class ElmanRNN(RNN):
"""
An Elman RNN that, at each time step,
takes vector input.
>>> e = ElmanRNN(5, 10)
>>> print(e.step(np.array([1, 2, 3, 4, 5])))
[0.53239282 0.50239672 0.50754092 0.49467159 0.47343499 0.48537873
0.48764785 0.50135803 0.51225634 0.49526006]
"""
def __init__(self,
input_size: int,
hidden_size: int) -> None:
super().__init__()
self.hidden_size = hidden_size
# in this vector version, weight matrices are actual
# matrices, and biases are vectors
self.params["Wh"] = np.random.randn(input_size, hidden_size) * 0.01
self.params["Uh"] = np.random.randn(hidden_size, hidden_size) * 0.01
self.params["bh"] = np.zeros(hidden_size,)
self.reset_state()
def reset_state(self) -> None:
self.hidden_state = np.zeros(self.hidden_size,)
def step(self, input_: np.ndarray) -> np.ndarray:
input_linear = np.dot(input_, self.params["Wh"])
hidden_linear = np.dot(self.params["Uh"], self.hidden_state) + self.params["bh"]
# compute new hidden state
self.hidden_state = sigmoid(input_linear + hidden_linear)
return self.hidden_state
def forward_sequence(self, sequence: Sequence[np.ndarray]) -> Sequence[np.ndarray]:
self.reset_state()
hidden_states = []
for item in sequence:
hidden_states.append(self.step(item))
return hidden_states
LSTMState = namedtuple("LSTMState", ["cell_states", "hidden_states"])
class LSTM(RNN):
"""
Long-short term memory RNN, vanilla without peephole
connections.
>>> e = LSTM(5, 10)
>>> print(e.step(np.array([1, 2, 3, 4, 5])))
[ 9.81669866e-05 3.49891068e-03 2.03026849e-02 3.67467230e-03
-5.13049272e-03 -4.68873686e-03 8.13246236e-03 5.22321287e-03
2.76350076e-03 -4.19133057e-03]
"""
def __init__(self,
input_size: int,
hidden_size: int) -> None:
super().__init__()
self.hidden_size = hidden_size
# forget gate parameters
self.params["Wf"] = np.random.randn(hidden_size, input_size) * 0.01
self.params["Uf"] = np.random.randn(hidden_size, hidden_size) * 0.01
self.params["bf"] = np.zeros(hidden_size,)
# input gate parameters
self.params["Wi"] = np.random.randn(hidden_size, input_size) * 0.01
self.params["Ui"] = np.random.randn(hidden_size, hidden_size) * 0.01
self.params["bi"] = np.zeros(hidden_size,)
# output gate parameters
self.params["Wo"] = np.random.randn(hidden_size, input_size) * 0.01
self.params["Uo"] = np.random.randn(hidden_size, hidden_size) * 0.01
self.params["bo"] = np.zeros(hidden_size,)
# cell state parameters
self.params["Wc"] = np.random.randn(hidden_size, input_size) * 0.01
self.params["Uc"] = np.random.randn(hidden_size, hidden_size) * 0.01
self.params["bc"] = np.zeros(hidden_size,)
self.reset_state()
def reset_state(self) -> None:
self.hidden_state = np.zeros(self.hidden_size,)
self.cell_state = np.zeros(self.hidden_size,)
def step(self, input_: np.ndarray) -> np.ndarray:
# forget gate
forget_gate_input = np.dot(self.params["Wf"], input_)
forget_gate_hidden = np.dot(self.params["Uf"], self.hidden_state)
forget_gate = sigmoid(forget_gate_input + forget_gate_hidden + self.params["bf"])
# input gate
input_gate_input = np.dot(self.params["Wi"], input_)
input_gate_hidden = np.dot(self.params["Ui"], self.hidden_state)
input_gate = sigmoid(input_gate_input + input_gate_hidden + self.params["bi"])
# output gate
output_gate_input = np.dot(self.params["Wo"], input_)
output_gate_hidden = np.dot(self.params["Uo"], self.hidden_state)
output_gate = sigmoid(output_gate_input + output_gate_hidden + self.params["bo"])
# cell state
cell_state_input = np.dot(self.params["Wc"], input_)
cell_state_hidden = np.dot(self.params["Uc"], self.hidden_state)
cell_state = tanh(cell_state_input + cell_state_hidden + self.params["bc"])
self.cell_state = (forget_gate * self.cell_state) + (input_gate * cell_state)
self.hidden_state = output_gate * tanh(self.cell_state)
return self.hidden_state, self.cell_state
def forward_sequence(self, sequence: Sequence[np.ndarray]) -> LSTMState:
self.reset_state()
hidden_states = []
cell_states = []
for item in sequence:
hidden_state, cell_state = self.step(item)
hidden_states.append(hidden_state)
cell_states.append(cell_state)
return LSTMState(hidden_states, cell_states)
class GRU(RNN):
"""
Gated recurrent unit RNN.
>>> g = GRU(5, 10)
>>> print(g.step(np.array([1, 2, 3, 4, 5])))
[-0.02601452 -0.00179476 -0.01031278 -0.02597636 -0.02946662 0.02102406
-0.04205244 -0.03658039 0.00968579 0.02418459]
"""
def __init__(self,
input_size: int,
hidden_size: int) -> None:
super().__init__()
self.hidden_size = hidden_size
# update gate parameters
self.params["Wz"] = np.random.randn(hidden_size, input_size) * 0.01
self.params["Uz"] = np.random.randn(hidden_size, hidden_size) * 0.01
self.params["bz"] = np.zeros(hidden_size,)
# reset gate parameters
self.params["Wr"] = np.random.randn(hidden_size, input_size) * 0.01
self.params["Ur"] = np.random.randn(hidden_size, hidden_size) * 0.01
self.params["br"] = np.zeros(hidden_size,)
# hidden proposal parameters
self.params["Wh"] = np.random.randn(hidden_size, input_size) * 0.01
self.params["Uh"] = np.random.randn(hidden_size, hidden_size) * 0.01
self.params["bh"] = np.zeros(hidden_size,)
self.reset_state()
def reset_state(self) -> None:
self.hidden_state = np.zeros(self.hidden_size,)
def step(self, input_: np.ndarray) -> np.ndarray:
# update gate
update_gate_input = np.dot(self.params["Wz"], input_)
update_gate_hidden = np.dot(self.params["Uz"], self.hidden_state)
update_gate = sigmoid(update_gate_input + update_gate_hidden + self.params["bz"])
# reset gate
reset_gate_input = np.dot(self.params["Wr"], input_)
reset_gate_hidden = np.dot(self.params["Ur"], self.hidden_state)
reset_gate = sigmoid(reset_gate_input + reset_gate_hidden + self.params["br"])
# hidden state proposal
proposal_input = np.dot(self.params["Wh"], input_)
proposal_hidden = np.dot(self.params["Uh"], (self.hidden_state * reset_gate))
proposal = tanh(proposal_input + proposal_hidden + self.params["bh"])
# new hidden state
self.hidden_state = ((1 - update_gate) * proposal) + (update_gate * self.hidden_state)
return self.hidden_state
def forward_sequence(self, sequence: Sequence[np.ndarray]) -> Sequence[np.ndarray]:
self.reset_state()
hidden_states = []
for item in sequence:
hidden_states.append(self.step(item))
return hidden_states